10 posts tagged with “ai-agent-security”
An open-source AI pentest platform that now promises "governed execution" turned up on infrastructure tied to more than 600 compromised FortiGate devices. The governing is done by whoever runs it, and that's true of your agents too.
A fake-font credential stealer landed in 12,137 GitHub repos between October 14 and 16, about 130 times the cluster in the only write-up so far. It goes straight for Claude Code, Gemini and OpenAI credentials. We found malware in nine forks of REA.
Between October 1 and 8, a proposed Senate bill, the Wikimedia Foundation and the UK's data regulator all put AI agent operators on notice. Whoever runs an agent answers for what it does to everyone else.
Agent control is turning into a product you can buy. Open or closed, you didn't write it, so what it does in your environment is the only proof you'll get.
Darktrace rewrote the saved conversations of four coding agents and got them to attack a network. The lesson isn't about signatures. It's about where your agent's permission comes from.
Seven defects surfaced in one migration. None of them crashed, none failed a test, and none turned a dashboard red. A broken control and a working control send the same signal: silence.
An agent refused to run the attacker's binary, then wrote its own decoder and ran the payload anyway. Three incidents and one of our own show the same thing: a constraint you hand an agent is a preference it can route around.
A tidy-up PR removed a safety pattern from a Snowflake workflow. Five days later an autonomous agent had a Jira token. The headlines said an AI wrote the bug; the commit history cannot actually tell you who did, and that is the more serious problem.
Five CVEs across four vendors in thirteen months, all the same bug: your coding agent trusts a directory, then executes a file that can change underneath it. Here is the pattern, and the one line in Claude Code's restricted mode that actually fixes it.
Three pieces of research showed how to hijack AI coding agents. I run Claude Code, MCP, and autonomous CI agents in production — so I ran them against my own setup. Here's what held, what didn't, and what I changed.